Agentic Coding on Coding unseen tasks (test)
29.2Pass@1SGE
Evaluation results
| Method | Links | |
|---|---|---|
| SGERL Training=Strategy-Guided Exploration (SGE), Checkpoints=1k updates, Aggregation=Average ± standard error across 3 random seeds2026.03 | 29.2 | |
| GRPORL Training=GRPO, Checkpoints=1k updates, Aggregation=Average ± standard error across 3 random seeds2026.03 | 22 | |
| Zero-ShotEvaluation Protocol=Direct base-LLM evaluation, RL Training=None2026.03 | 13.5 |